Automatic Open Space Area Extraction and Change Detection from High Resolution Urban Satellite Images

Transcription

1 Automatic Open Space Area Extraction and Change Detection from High Resolution Urban Satellite Images Dr. Deshmukh Nilesh Kailasrao School of Computational Sciences, Swami Ramanand Teerth Marathwada University, Nanded, (MH), India. ABSTRACT Timely and accurate change detection of Earth's surface features is extremely important for understanding relationships and interactions between human and natural phenomena in order to promote better decision making. Remote sensing data are primary sources extensively used for change detection in recent decades. Many change detection techniques have been developed. This research paper presents an efficient and reliable automatic extraction algorithm to find out the open space area from the high resolution urban satellite imagery, and to detect changes from the extracted open space. This research paper depicts the efficient and reliable automatic extraction algorithm to find out the open space from the high resolution urban satellite imagery, and detecting changes from the extracted open space during the period 2003, 2006 and This automatic extraction and change detection algorithm makes use of some filters, segmentations and grouping and optimization techniques applied on satellite images. The resultant images may be used to calculate the total available open area and the built up area. It may also be used to compare the difference between present and past open space using historical urban satellite images of that same projection. Key words: Change detection, Spatial database, Geospatial tools, Spatial information system, Open source software. Corresponding Author: Dr. Deshmukh N.K. INTRODUCTION Extraction of open space area from raster images is a very important part of GIS features such as GIS updating, geo-referencing and geo spatial data integration. However extracting open space area from raster image is a time consuming operation when performed manually, especially when the image is complex. The automatic extraction of open space area is critical and essential to the fast and effective processing of large number of raster images in various formats, complexities and conditions. How open space area are extracted properly from raster images depend on how open space area appear in raster image. In this research paper, we study automatic extraction of open space area from high resolution urban area satellite image. A high resolution satellite image typically has a resolution of 0.5 to 1.0 m. Under such high resolution, an open space is not same any more in whole image. Instead, objects such as lake(s), trees are easily identifiable. This class of images contains very rich information and when fused with vector map can provide a comprehensive view of a geographical area. Google, Yahoo, and Virtual Earth maps Page 130

2 are good examples to demonstrate the power of such high resolution images [1]. However, high resolution images pose great challenges for automatic feature extraction due to the inherent complexities. First, a typical aerial photo captures everything in the area such as buildings, cars, trees, open space area, etc. Second, different objects are not isolated, but mixed and interfere with each other, e.g., the shadows of trees on the road, building tops with similar materials. Third, roads may even look like open space area, which have the same characteristics, and hence fail to extract total open space area. In addition, the light and weather conditions have big impact over images. Therefore, it is impossible to predict what and where objects are, and how they look like in a raster image. All these uncertainties and complexities make the extraction very difficult. Due to its importance, much effort has been devoted to this problem [2] [3]. Unfortunately, there are still no existing methods that can deal with all these problems effectively and reliably. Some typical high resolution images are shown in Fig.1 and they show huge difference among them in terms of the color spectrum and noise level [4]. There are numerous factors that can distort the edges, including but not limited to blocking objects such as trees and shadows, surrounding objects in similar colors such as roof tops. As a matter of fact, the result of edge detection is as complicated as the image itself. Edges of open space area are either missing or broken and straight edges correspond to buildings, as shown in Fig.2. Therefore, edge-based extraction schemes fail to produce reliable results under such circumstances. (a) (b) (c) Fig.1: Examples of high resolution images, (a) An aerial photo for an area in the Latur city, (b) High resolution urban satellite image. (Latur city, dated 23 Feb. 2003), (c) Google map satellite image. Page 131

3 METHODOLOGY In this research paper, we develop an integrated scheme for automatic extraction that exploits the inherent nature of open space area. Instead of relying on the edges to detect open space, it tries to find the pixels that belong to the same open area region based on how they are related visually and geometrically. Studies have shown that the visual characteristic of a open space is heavily influenced by its physical characteristics such as material, surface condition. It is impossible to define a common pattern just based on color or spectrum to uniquely identify the open space area. In our scheme, we consider a open space as a group of similar pixels. The similarity is defined in the overall shape of the region they belong to, the spectrum they share, and the geometric property of the region. Different from edge-based extraction schemes, the new scheme first examines the visual and geometric properties of pixels using a new method. The pixels are identified to represent each region. All the regions are verified based on the general visual and geometric constraints associated with a open space area. Therefore, the roof top of a building or a long strip of trees is not misidentified as a open space area segment. There is also no need to assume or guess the color spectrum of open space area, which varies greatly from image to image according to the example images. Fig.2: Edge extraction (using canny method) of Fig.1 (b). As illustrated by examples in Fig.1, an open space area is not always a contiguous region of regular linear shape, but a series of segments with no constant shapes. This is because each segment may include pixels of surrounding objects in similar colors or miss some of its pixels due to interference of surrounding objects. A reliable extraction scheme must be able to deal with such issues. In the following sections, we discuss how to capture the essence of similarity, translate them and finally turn them into display. IDENTIFICATION OF WATER BODIES FROM SATELLITE IMAGES While extracting open space area from high resolution urban satellite images, first, we should identify the water bodies like lakes or reservoirs or tanks or rivers etc. For this purpose, a large number of earth observation satellites has orbited, and is orbiting our planet to provide frequent imagery of its surface. From these satellites, many can potentially provide useful Page 132

4 information for assessing erosion, although seldom it has actually been used for this purpose. This section provides a brief overview of the space borne sensors applied in water-body extraction studies. Optical satellite systems have most frequently been applied in water body extraction research. The parts of the electromagnetic spectrum covered by these sensors include the visible and near infrared (VNIR) ranging from 0.4 to 1.3 μm, the shortwave infrared (SWIR) between 1.3 and 3.0 μm, the thermal infrared (TIR) from 3.0 to 15.0 μm and the long-wavelength infrared (LWIR) from (7-14 μm). The Table 1 summarizes sensor characteristics of the systems used [5]. Table 1: The optical satellite sensors applied for water body extraction Satellites Sensors Operation Time Spatial Resolution Landsat-1,2,3 MSS m NOAA/TIROS AVHRR 1978-Present 1001m Nimbus-7 CZCS m Landsat-4,5 TM m SPOT-1,2,3 HRV Present 20m IRS- 1A,1B LISS-1 LISS m 36.25m IRS-1C,1D PAN 1995-Present 5.8m SPOT-4 HRVIR 1998-Present 10m IKONOS Panchromatic Multispectral 1999-Present 1m 2m Landsat-7 ETM 1999-Present 15m 30m TERRA ASTER MODIS 1999-Present 1999-Present 15m 250m QuickBird Panchromatic Multispectral 2001-Present 0.61m SPOT-5 Panchromatic Multispectral 2002-Present 5m WorldView-1 Panchromatic 2007-Present 0.55m GEOEYE-1 Pan-sharpened Panchromatic Multispectral 2008-Present 0.41m AUTOMATIC EXTRACTION SCHEME The goal of the proposed system is to develop a complete and practical automatic solution to the extraction of open space area problem for high resolution aerial/satellite images. The first stage extracts open space areas that are relatively easier to identify such as major grounds and the second stage deals with open space that are harder to identify. The reason for Page 133

5 such design is a balance between reliability and efficiency. Some open space areas are easier to identify because they are more identifiable and contain relatively less noise. Since some open space area in the same image share some common visual characteristics, the information from the already extracted area and other objects, such as spectrum, can be used to simplify the process of identifying open space area that are less visible or heavily impacted by surrounding objects or by different colors. Otherwise, such areas are not easily distinguishable from patterns formed by other objects. For example, a set of collinear blocks may correspond to a open space or a group of buildings (houses) from the same block. The second stage also serves an important purpose to fill the big gaps for open space extracted in stage one. Under some severe noise, part of the area may be disqualified as valid open space region and hence missed in stage one, leaving some major gaps in open space area edges. With the additional spectrum information, these missed areas can be easily identified to complete the open space area extraction. Therefore, the two stage process eliminates the need to assume or guess the color spectrum of open space and allow them to extract much more complete open area. Fig.3: Automatic extraction scheme. Each major stage consists of the following three major steps: filtering, grouping and optimization, as shown in Fig.3. The details of each step will be discussed in the following section. ALGORITHM AND RESULTS Here, we assume images to satisfy the following two general assumptions. These two assumptions are derived based on the minimum conditions for open space area to be identifiable and therefore are easily met by most images. Visual constraint: majority of the pixels from the same open space area have similar spectrum that is distinguishable from most of the surrounding areas; Page 134

6 Geometric constraint: a open space is a region that is relatively has no standard shape, compared with other objects in the image; These two constraints are different from the assumption. The visual constraint does not require a open space region to have single color or constant intensity. It only requires blank area to look visually different from surrounding objects in most parts. The geometric constraint does not require a smooth edge, only the overall shape to be a long narrow strip. So these conditions are much weaker and a lot more practical. As we can see, these assumptions can accommodate all the difficult issues very well, including blurring, broken or missing edge of open area boundaries, heavy shadows, and interfering surrounding objects [6]. A. Filtering The step of filtering is to identify the key pixels that will help determine if the region they belong to is likely an open space area segment. Based on the assumption of visual constraint, it is possible to establish an image segmentation using methods of edge detection. We notice that such separation of regions is required to be precise and normally contains quite a lot of noise. In the best, the boundaries between regions are a set of line segments for most images, as is the case shown in Fig.2. It certainly does not tell which region corresponds to which area and which is not based on the extracted edges. As a matter of fact, most of the regions are not completely separated by edges and are still interconnected based on 4-connect path or 8- connect path. In order to fully identify and separate open space regions from the rest of image, we have proposed to invert them and again extract the edges using Sobel edge detector to highlight sharp changes in intensity in the active image or selection. Two 3x3 convolution kernels (shown below) are used to generate vertical and horizontal derivatives. The final image is produced by combining the two derivatives using the square root of the sum of the squares [7]. Fig.4: Result after edge detection & outlier removal of Fig.2 Page 135

7 The next step is a removal of outliers through a process in which we can replace a pixel by the median of the pixels in the surrounding if it deviates from the median by more than a certain value (the threshold).we used the following values for outlier removal: R = 10.0 pixels, T = 2, OL = Bright R is radius that determines the area used for calculating the median (uncalibrated, i.e., in pixels). The Fig.4 shows how radius translates into an area. The T (Threshold) determinates by how much the pixel must deviate from the median to get replaced, in raw (uncalibrated) units. Which Outlier OL determines whether pixels brighter or darker than the surrounding (the median) are replaced. B. Segmentation The step of segmentation is to verify which region is a possible road region based on the central pixels. Central pixels contain not only the centerline information of a region, but also the information of its overall geometric shape. For example, a perfect square will only have one central pixel at its center. A long narrow strip region will have large number of central pixels. Only regions with ratios above certain thresholds are considered to be candidate regions. In order to filter out interference as much as possible for reliable extraction during the first major stage, a minimum region width can be imposed. This will effectively remove most of the random objects from the image. However, such width constraint will be removed during the second major step as improper regions can also be filtered out based on the color spectrum information obtained from stage one. Therefore, small regions with close spectrum are examined for possible small open space area in the second stage [8]. In addition to the geometry constraint, some general visual information can also be applied to filter out obvious non-open space regions. For example, if the image is a color image, most of the tree and grass areas are greenish. Also tree areas usually contain much richer textures than normal smooth road surfaces. The intensity transformation and spatial and frequency filtering can be used to filter out such areas. The minimum region assumption of the proposed scheme does not exclude the use of additional visual information to further improve the quality of extraction if they are available. C. Grouping and optimization As the result of segmentation, the open space area segments are typically incomplete and disconnected due to heavy noise. Page 136

8 Fig.5: Result after applying upper & lower threshold values. The purpose of this step is to group corresponding open space area segments together in order to find the optimal results for required area extraction. If enough information is available to determine the open space area spectrum, then optimization is better applied after all the segments are identified. The Fig.5 is the result of threshold used to automatically or interactively set lower=0 and upper=48 threshold values, segmenting the image into features of interest and background. The threshold features are displayed in white and background is displayed in red color i.e. the total open space area in the given projected satellite image is shown in Fig.6. (a) (b) Fig.6: Extracted open space area using historical images. (a) Open space area on Feb , (b) Open space area on Jan Page 137

9 The extracted open space area from high resolution urban satellite imagery is shown in Fig.7 with region wise numbers. The total available open space region s labels, area and centroids are calculated and shown in Table 2. Fig.7. Labeled regions of extracted open space area (Feb. 2003) Table 2. Area and centroids of labeled regions of Fig.7. Region Region Centroid of region labels area x1 y Page 138

10 Fig.8. Labeled regions of extracted open space area (Feb. 2006). The Table 3. shows the details of calculated area and centroids x1, y1 of labeled regions of open space of Fig.8. Table 3. Area and centroids of labeled regions of open space of Fig.8. Region Region Centroid of region labels area x1 y Page 139

11 Fig.9. Labeled regions of extracted open space area (Jan 2008). The Table 4 shows the details of calculated area and centroids x1, y1 of labeled regions of open space of fig.8. Table 4. Area and centroids of labeled regions of open space of Fig.9. Region Region Centroid of region labels area x1 y Based on the extracted areas from Table 1, 2, and 3, the comparison of the open space area from existing historical images of year 2003, 2006 and 2008 with respect to selected regions (Fig.7) is shown in the Fig.10. Page 140

12 Area INTERNATIONAL JOURNAL OF COMPUTER APPLICATION Labeled regions Fig.10. Comparative results of extracted open space areas of selected regions of imageries of year 2003, 2006 and CHANGE DETECTION Timely and accurate change detection of Earth's surface features is extremely important for understanding relationships and interactions between human and natural phenomena in order to promote better decision making. Remote sensing data are primary sources extensively used for change detection in recent decades [9]. Change detection is used to highlight or identify significant differences in imagery acquired at different times. It plays an important role in the lifecycle of GIS [10]. Based on the results of above experiment, the Fig.11, 12 and 13 show the changes detected during the period 2003 to 2008, 2003 to 2006 and 2006 to 2008 obtained by the corresponding year s satellite images respectively. Fig.11. Changes detected during 2003 to 2008 Fig.12. Changes detected during 2003 to Page 141

13 CONCLUSION Fig.13. Changes detected during 2006 to In this research paper, we have proposed a new automatic system for extracting open space area and change detection during a certain time period. The proposed approach is efficient, reliable, and assumes no prior knowledge about the required open space area, conditions and surrounding objects can be identified using different techniques to determine whether the region belongs to open space are or not?. It is able to process complicated aerial/satellite images from a variety of sources including aerial photos from Google and Yahoo satellite images. The applications of the proposed study are: the emergency landing of the helicopters in open space area, urban development planning, land use/land cover assessment, natural disaster area assessment, etc. The experimental results have demonstrated its efficacy. REFERENCE [1] Anji Reddy M, Text book of Remote Sensing and Geographical Information System, 3rd ed., B.S. Publications, Hyderabad (India), [2] J.B. Mena. State of the Art on Automatic Road Extraction for GIS Update: a Novel Classification. Pattern Recognition Letters, 24(16), 2003, pp [3] M.F. Auclair-Fortier, D. Ziou, C. Armenakis, and S. Wang. Survey of Work on Road Extraction in Aerial and Satellite Images. Technical Report 241, Departement de mathématiques et d informatique, Université de Sherbrooke, [4] Gorokhovich, and A. Voustianiouk, Remote Sensing of Environment, Center for International Earth Science, 104, 2006, pp [5] Rajiv Kumar Nath, S K Deb, Waterbody area extraction from high resolution satellite images, IJIP, Vol. 3, Issue 6, 2009, pp [6] Y. Li and R. Briggs, Scalable and error tolerant automated georeferencing under affine transformations. IEEE International Geoscience and Remote Sensing Symposium, Boston, MA, July [7] Rafeal C. Gonzalez and Richard E. Woods, Digital Image Processing, 2nd ed., Prentice- Hall, New Jersey, [8] Anil K. Jain, Fundamentals of Image Processing, Prentice-Hall, New Jersey, Page 142

14 [9] D. Lu, P. Mausel, E. Brondízio, E. Moran, Change Detection Techniques, International Journal of Remote Sensing, Volume 25, Number 12, 2004, pp [10] Ashbindu Singh, Review Article Digital change detection techniques using remotelysensed data, International Journal of Remote Sensing, Volume: 10, Issue: 6, Publisher: Taylor & Francis,1989, pp Page 143